Abstract:
Short text clustering (STC) is an important task that can discover topics or groups in the fast-growing social networks, e.g., Tweets and Google News. Different from the long texts, STC is more challenging since the word co-occurrence patterns presented in short texts usually make the traditional methods (e.g., TF-IDF) suffer from a sparsity problem of inevitably generating sparse representations. Moreover, these learned representations may lead to the inferior performance of clustering which essentially relies on calculating the distances between the presentations. For alleviating this problem, recent studies are mostly committed to developing representation learning approaches to learn compact low-dimensional embeddings, while most of them, including probabilistic graph models and word embedding models, require all documents in the corpus to be present during the training process. Thus, these methods inherently perform transductive learning which naturally cannot handle well the representations of unseen documents where few words have been learned before. Recently, Graph Neural Networks (GNNs) has drawn a lot of attention in various applications. Inspired by the mechanism of vertex information propagation guided by the graph structure in GNNs, we propose an inductive document representation learning model, called IDRL, that can map the short text structures into a graph network and recursively aggregate the neighbor information of the words in the unseen documents. Then, we can reconstruct the representations of the previously unseen short texts with the limited numbers of word embeddings learned before. Experimental results show that our proposed method can learn more discriminative representations in terms of inductive classification tasks and achieve better clustering performance than state-of-the-art models on four real-world datasets.
